Machine Learning with Swift

BNNS example

In the following example, input images are of size 224 x 224 x 64 and output images are of size 222 x 222 x 96. The dimensionality of convolution weights is 3 x 3 x 64 x 96. That's 5.45 billion floating-point operations (gigaFloPS). In a whole MNIST recognition network, it's about 1-2 trillion operations per forward pass.

BNNS is a part of Accelerate, so you need to import Accelerate to access the neural networks building blocks. The first thing you do is describing the input stack:

var inputStack = BNNSImageStackDescriptor(  width: 224, height: 224, channels: 64,  row_stride: 224, image_stride: 224*224,  data_type: BNNSDataTypeFloat32,  data_scale: 1.0, data_bias: 0.0)

Most of the parameters are self-evident; row_stride is an increment ...